Decoded Traffic as Free Pilots: Tag-Heralded Self-Interference Calibration for Full-Duplex Dynamic Relaying

A relay that helps a source reach a distant destination would ideally listen and talk at the same time on the same frequency. This is called full-duplex operation. The obstacle is self-interference: the relay's own transmission leaks into its receiver at 50 to 110 dB above the signal it is trying to hear. Cancelling that leakage digitally requires a model of the relay's own nonlinear transmit chain, and that model drifts. It is normally refit using calibration pilots, which are known symbols that cost airtime. This proposal observes that a decoding relay already produces something better than pilots. Once it decodes the source's packet and the error-detecting checksum (CRC) confirms the decode, the relay knows exactly what the source sent. It has always known exactly what it sent itself. Subtract the reconstructed source contribution from the buffered received samples, and what remains is attributable to self-interference plus noise. Every successfully decoded packet therefore becomes a full-length, verified training sequence for the self-interference canceller, at zero airtime cost.

The idea borrows its logic from double-tag reconstruction in particle physics. There, fully reconstructing one particle of a threshold-produced pair, the "tag," fixes the partner's four-momentum by conservation. The unobserved remainder, such as a neutrino, is then measured as a "missing" quantity rather than modelled. Ratios of tagged yields also cancel unknown normalizations such as luminosity (Absolute Measurement of Hadronic Branching Fractions of the Ds+ Meson; Measurement of cross sections for D0D¯0 and D+D− production at √s=3.773 GeV). We use this as a design principle, not a borrowed technique. The verified decode is the tag. Self-interference is the missing quantity. A ratio of residual energies measured on tagged windows plays the role of the luminosity-cancelling ratio and removes unknown front-end gain. We embed this calibration loop in a dynamic decode-and-forward relay. That is a relay that listens until it has gathered enough information to decode and then switches to transmitting. We let it choose among three states: listen-only, full-duplex, and transmit-only. Its switching rule accounts for the residual self-interference its own canceller leaves behind.

Core Claim

Our claim has three parts. First, a full-duplex dynamic decode-and-forward relay can be trained only on CRC-verified decoded traffic, with no dedicated pilots. Such a relay reaches the residual self-interference floor of a pilot-trained canceller. Second, it beats half-duplex dynamic decode-and-forward in outage at multiplexing gains above 0.5, the regime where half-duplex DDF is known to fall short of the genie-aided bound. Third, the switching rule must be "interference-aware": it measures accumulated information with the post-cancellation interference in the denominator. Ignoring this either overstates the full-duplex benefit or causes decode failures. Those failures starve the canceller of tags and create a self-reinforcing collapse.

The new element is the closed loop between information-driven switching and tag-driven calibration. Switching determines how many tags arrive. Tags determine canceller quality. Canceller quality determines how fast the relay accumulates information, and that in turn drives switching.

Positioning Against Prior Work

The relaying side starts from On the achievable diversity-multiplexing tradeoff in half-duplex cooperative channels. That paper shows dynamic decode-and-forward attains the genie-aided diversity–multiplexing tradeoff $d(r) = (1-r) + (1-2r)$ for $0 \le r \le 0.5$. The tradeoff says how fast outage falls with SNR at a given rate growth. Above $r = 0.5$ the relay spends more than half the block decoding, and this motivates our high-rate target. Diversity-Multiplexing Tradeoffs in MIMO Relay Channels compares half-duplex DDF against full-duplex relaying, but treats full duplex as an idealized upper bound with perfect cancellation. We replace that ideal with a measured, learned residual. When are dynamic relaying strategies necessary in half-duplex wireless networks? and Generalized Diversity-Multiplexing Tradeoff of Half-Duplex Relay Networks map where dynamic schedules beat static ones. Both stay strictly half-duplex, so they supply the switching vocabulary but say nothing about self-interference.

On the hardware side, Full-Duplex MIMO Relaying: Achievable Rates under Limited Dynamic Range models transceiver distortion that scales with signal power. It identifies a "critical INR" (interference-to-noise ratio) above which full duplex should back off toward half duplex. This constrains our design: no learned canceller removes distortion created in a saturated front-end, so our relay must be able to fall back. That paper's back-off is a static optimization, whereas our fall-back is decided per block from accumulated information. Sample-Efficient Self-Interference Cancellation for In-Band Full Duplex Radios via In-Context Learning supplies the canceller we reuse. It is a GPT-2-style transformer that treats calibration samples as context and predicts nonlinear residual self-interference. It reaches its floor with about 64× fewer calibration samples than model-based baselines on USRP N210 hardware. It also reports that performance drops by roughly 10 dB on a new device without gain normalization. That paper still feeds its model dedicated pilots on a single point-to-point link. We feed it decoded traffic inside a relay protocol whose transmit bursts start at random times.

The closest conceptual neighbour is Self-interference cancellation in multi-hop full-duplex networks via structured signaling. It estimates self-interference on the fly by having the source leave signal levels empty, so the relay's own leakage can be observed there. It works in a deterministic bit-pipe model and requires the source to sacrifice rate. Our scheme sacrifices nothing at the source. The "probe" is the full data packet, made known after the fact by a verified decode, in a Gaussian fading model with real nonlinear hardware.

Research Program

The relay buffers its post-analog-cancellation digital samples for each block. When a decode passes the CRC, it re-encodes the source codeword and reconstructs the source contribution using its channel estimate. It subtracts this from the buffer, leaving a "recoil" sequence of self-interference plus noise, paired with its own known transmit samples. These pairs are appended to the in-context canceller's context window, and stale pairs are evicted to track drift.

To remove device-dependent gain without a separate normalization stage, we form a ratio across tagged windows. The ratio is residual energy during full-duplex segments divided by residual energy during listen-only segments within the same block. This mirrors the way the physics ratio of double-tag to single-tag yields cancels the unknown pair count.

The switching rule generalizes DDF's accumulated-information test. The relay commits to decoding at the first normalized time $t_0$ satisfying

$$\frac{1}{T}\sum_{\tau \le t_0 T} \log\!\left(1 + \frac{\mathrm{SNR}\,|h_{SR}|^2}{1 + \mathrm{INR}_{\mathrm{res}}(\tau)}\right) \ge R .$$

Here $T$ is the block length and $R$ is the target rate. $h_{SR}$ is the source–relay channel. $\mathrm{INR}_{\mathrm{res}}(\tau)$ is the residual interference-to-noise ratio after cancellation at time $\tau$. That residual is zero during listen-only segments and equals the canceller's tracked residual during segments where the relay is also forwarding the previous block. Before each block, the relay compares the predicted $t_0$ under listen-only and under full-duplex operation, and picks the mode with the larger expected forwarding time. It falls back to listen-only whenever the predicted residual exceeds the critical-INR regime of Day et al.

On the analysis side, we will derive the tradeoff curve of this protocol under a residual that scales as $\mathrm{INR}_{\mathrm{res}} \propto \mathrm{SNR}^{\beta}$ for $\beta < 1$. That scaling is the form produced when estimation error from $n$ tagged samples shrinks like $1/n$ while the leakage grows with transmit power. We will also show how the limited-dynamic-range ceiling caps $\beta$ from below.

Baselines

The naive reference point is half-duplex dynamic decode-and-forward exactly as in Azarian et al., with no full-duplex hardware at all. It defines the bar the whole effort must clear, especially above $r = 0.5$, and for $r \le 0.5$ it is already optimal, so we expect only to match it there.

The strongest existing competitor is a full-duplex dynamic relay using the in-context canceller of Balaji et al. trained on dedicated pilots. We include it at matched pilot budgets of 0%, 2%, 5% and 10% of airtime, with their gain normalization. The best reported number for that method is the 64× sample-efficiency advantage over model-based baselines on USRP N210 captures. On the synthetic power-amplifier benchmark it is 4.4× more sample-efficient than widely-linear alternating least squares. No reported end-to-end relay outage figure exists on Lacuna for any full-duplex dynamic decode-and-forward scheme, so we will produce that number ourselves for both systems. Beating this baseline means one of two things. Either we match its residual self-interference within 1 dB and its outage curve while it spends at least 5% airtime on pilots and we spend none. Or we strictly beat it at equal net spectral efficiency once pilot overhead is charged.

The third baseline is the static full-duplex decode-and-forward relay with power back-off of Day et al., given the same canceller. This isolates the value of the dynamic switching rule.

We include three ablations. The first replaces CRC-verified tags with unverified hard decisions, ordinary decision-directed estimation. This isolates whether the exactness of the tag matters. The second removes the residual ratio and uses the paper's gain normalization instead, isolating the luminosity-style ratio. The third uses a switching rule that ignores $\mathrm{INR}_{\mathrm{res}}$. All systems share the same captures, channel draws, canceller architecture and parameter count, and the same hyperparameter search budget.

Why not just…?

Why not just keep half-duplex DDF? Below $r = 0.5$ it is already optimal, and we concede that regime. Above it, its diversity falls off because decoding consumes more than half the block (Azarian et al.). A deep source–relay fade also leaves the relay silent for the whole block. The outage-versus-SNR slope at $r = 0.6$ to $0.8$ in our experiments settles whether full duplex buys anything once real residual interference is charged.

Why not just use periodic pilots with the in-context canceller? This is the real threat, because that canceller is already very sample-efficient. Its 64× advantage means pilot overhead could be small enough that our savings are marginal. The pilot-budget sweep answers this directly. If a 1% pilot budget already matches tag-trained performance under realistic drift, our contribution shrinks to an overhead saving. We will report it as such. Our expectation that tags win rests on volume and timing rather than sample efficiency. Tags supply thousands of samples per block, captured exactly in the transient right after the relay starts transmitting, which pilots placed at fixed times miss.

Why not just do decision-directed estimation, which every receiver already does? It is decision-directed estimation, but with two differences. CRC verification makes the tag exact rather than error-prone. More importantly, a wrong decision would corrupt the context of a learned canceller, and a wrong canceller causes further decode failures. This failure path is specific to the closed loop. The unverified-decision ablation measures it. If the ablation performs as well as verified tags, the CRC gating is unnecessary and we will say so.

Why not just use dynamic quantize-map-and-forward and avoid decoding? That scheme (Kolte et al.) avoids decoding delay. However, it produces no tag, so it has no free training signal. It also forwards whatever residual self-interference it quantizes. It is out of scope as a competitor but is a natural follow-up comparison.

Proposed Experiments

We build a semi-synthetic relay emulator. Measured self-interference comes from USRP N210 full-duplex captures collected with the protocol of Balaji et al.. Power-amplifier nonlinearity at wider bandwidth comes from the APA_200MHz GaN Doherty dataset of OpenDPDv2. Synthetic source signals pass through Rayleigh source–relay, relay–destination and source–destination channels at controlled SNR, so the true source contribution is known for scoring.

Metrics are residual self-interference NMSE, outage probability versus SNR at multiplexing gains from 0.3 to 0.8, the empirical diversity slope, net spectral efficiency after pilot overhead, and the rate of tag-starvation collapse under drift. The emulator imposes drift by switching between the three captured hardware units mid-run.

The decisive check has two parts. First, at $r \ge 0.6$, the tag-trained full-duplex dynamic relay must show a steeper outage slope than half-duplex DDF. Second, it must match the 5%-pilot in-context baseline at zero pilots. A second check is the cross-device transfer. The residual-ratio normalization should recover most of the roughly 10 dB loss Balaji et al. report without gain normalization. A third is a stress test at INR beyond the critical regime. There, the interference-aware rule should fall back to listen-only and track half-duplex DDF, while the naive rule should degrade below it.